1 # - Tools for building CUDA C files: libraries and build dependencies.
2 # This script locates the NVIDIA CUDA C tools. It should work on linux, windows,
3 # and mac and should be reasonably up to date with CUDA C releases.
5 # This script makes use of the standard find_package arguments of <VERSION>,
6 # REQUIRED and QUIET. CUDA_FOUND will report if an acceptable version of CUDA
9 # The script will prompt the user to specify CUDA_TOOLKIT_ROOT_DIR if the prefix
10 # cannot be determined by the location of nvcc in the system path and REQUIRED
11 # is specified to find_package(). To use a different installed version of the
12 # toolkit set the environment variable CUDA_BIN_PATH before running cmake
13 # (e.g. CUDA_BIN_PATH=/usr/local/cuda1.0 instead of the default /usr/local/cuda)
14 # or set CUDA_TOOLKIT_ROOT_DIR after configuring. If you change the value of
15 # CUDA_TOOLKIT_ROOT_DIR, various components that depend on the path will be
18 # It might be necessary to set CUDA_TOOLKIT_ROOT_DIR manually on certain
19 # platforms, or to use a cuda runtime not installed in the default location. In
20 # newer versions of the toolkit the cuda library is included with the graphics
21 # driver- be sure that the driver version matches what is needed by the cuda
24 # The following variables affect the behavior of the macros in the script (in
25 # alphebetical order). Note that any of these flags can be changed multiple
26 # times in the same directory before calling CUDA_ADD_EXECUTABLE,
27 # CUDA_ADD_LIBRARY, CUDA_COMPILE, CUDA_COMPILE_PTX or CUDA_WRAP_SRCS.
29 # CUDA_64_BIT_DEVICE_CODE (Default matches host bit size)
30 # -- Set to ON to compile for 64 bit device code, OFF for 32 bit device code.
31 # Note that making this different from the host code when generating object
32 # or C files from CUDA code just won't work, because size_t gets defined by
33 # nvcc in the generated source. If you compile to PTX and then load the
34 # file yourself, you can mix bit sizes between device and host.
36 # CUDA_ATTACH_VS_BUILD_RULE_TO_CUDA_FILE (Default ON)
37 # -- Set to ON if you want the custom build rule to be attached to the source
38 # file in Visual Studio. Turn OFF if you add the same cuda file to multiple
41 # This allows the user to build the target from the CUDA file; however, bad
42 # things can happen if the CUDA source file is added to multiple targets.
43 # When performing parallel builds it is possible for the custom build
44 # command to be run more than once and in parallel causing cryptic build
45 # errors. VS runs the rules for every source file in the target, and a
46 # source can have only one rule no matter how many projects it is added to.
47 # When the rule is run from multiple targets race conditions can occur on
48 # the generated file. Eventually everything will get built, but if the user
49 # is unaware of this behavior, there may be confusion. It would be nice if
50 # this script could detect the reuse of source files across multiple targets
51 # and turn the option off for the user, but no good solution could be found.
53 # CUDA_BUILD_CUBIN (Default OFF)
54 # -- Set to ON to enable and extra compilation pass with the -cubin option in
55 # Device mode. The output is parsed and register, shared memory usage is
56 # printed during build.
58 # CUDA_BUILD_EMULATION (Default OFF for device mode)
59 # -- Set to ON for Emulation mode. -D_DEVICEEMU is defined for CUDA C files
60 # when CUDA_BUILD_EMULATION is TRUE.
62 # CUDA_GENERATED_OUTPUT_DIR (Default CMAKE_CURRENT_BINARY_DIR)
63 # -- Set to the path you wish to have the generated files placed. If it is
64 # blank output files will be placed in CMAKE_CURRENT_BINARY_DIR.
65 # Intermediate files will always be placed in
66 # CMAKE_CURRENT_BINARY_DIR/CMakeFiles.
68 # CUDA_HOST_COMPILATION_CPP (Default ON)
69 # -- Set to OFF for C compilation of host code.
71 # CUDA_HOST_COMPILER (Default CMAKE_C_COMPILER, $(VCInstallDir)/bin for VS)
72 # -- Set the host compiler to be used by nvcc. Ignored if -ccbin or
73 # --compiler-bindir is already present in the CUDA_NVCC_FLAGS or
74 # CUDA_NVCC_FLAGS_<CONFIG> variables. For Visual Studio targets
75 # $(VCInstallDir)/bin is a special value that expands out to the path when
76 # the command is run from withing VS.
79 # CUDA_NVCC_FLAGS_<CONFIG>
80 # -- Additional NVCC command line arguments. NOTE: multiple arguments must be
81 # semi-colon delimited (e.g. --compiler-options;-Wall)
83 # CUDA_PROPAGATE_HOST_FLAGS (Default ON)
84 # -- Set to ON to propagate CMAKE_{C,CXX}_FLAGS and their configuration
85 # dependent counterparts (e.g. CMAKE_C_FLAGS_DEBUG) automatically to the
86 # host compiler through nvcc's -Xcompiler flag. This helps make the
87 # generated host code match the rest of the system better. Sometimes
88 # certain flags give nvcc problems, and this will help you turn the flag
89 # propagation off. This does not affect the flags supplied directly to nvcc
90 # via CUDA_NVCC_FLAGS or through the OPTION flags specified through
91 # CUDA_ADD_LIBRARY, CUDA_ADD_EXECUTABLE, or CUDA_WRAP_SRCS. Flags used for
92 # shared library compilation are not affected by this flag.
94 # CUDA_VERBOSE_BUILD (Default OFF)
95 # -- Set to ON to see all the commands used when building the CUDA file. When
96 # using a Makefile generator the value defaults to VERBOSE (run make
97 # VERBOSE=1 to see output), although setting CUDA_VERBOSE_BUILD to ON will
98 # always print the output.
100 # The script creates the following macros (in alphebetical order):
102 # CUDA_ADD_CUFFT_TO_TARGET( cuda_target )
103 # -- Adds the cufft library to the target (can be any target). Handles whether
104 # you are in emulation mode or not.
106 # CUDA_ADD_CUBLAS_TO_TARGET( cuda_target )
107 # -- Adds the cublas library to the target (can be any target). Handles
108 # whether you are in emulation mode or not.
110 # CUDA_ADD_EXECUTABLE( cuda_target file0 file1 ...
111 # [WIN32] [MACOSX_BUNDLE] [EXCLUDE_FROM_ALL] [OPTIONS ...] )
112 # -- Creates an executable "cuda_target" which is made up of the files
113 # specified. All of the non CUDA C files are compiled using the standard
114 # build rules specified by CMAKE and the cuda files are compiled to object
115 # files using nvcc and the host compiler. In addition CUDA_INCLUDE_DIRS is
116 # added automatically to include_directories(). Some standard CMake target
117 # calls can be used on the target after calling this macro
118 # (e.g. set_target_properties and target_link_libraries), but setting
119 # properties that adjust compilation flags will not affect code compiled by
120 # nvcc. Such flags should be modified before calling CUDA_ADD_EXECUTABLE,
121 # CUDA_ADD_LIBRARY or CUDA_WRAP_SRCS.
123 # CUDA_ADD_LIBRARY( cuda_target file0 file1 ...
124 # [STATIC | SHARED | MODULE] [EXCLUDE_FROM_ALL] [OPTIONS ...] )
125 # -- Same as CUDA_ADD_EXECUTABLE except that a library is created.
127 # CUDA_BUILD_CLEAN_TARGET()
128 # -- Creates a convience target that deletes all the dependency files
129 # generated. You should make clean after running this target to ensure the
130 # dependency files get regenerated.
132 # CUDA_COMPILE( generated_files file0 file1 ... [STATIC | SHARED | MODULE]
134 # -- Returns a list of generated files from the input source files to be used
135 # with ADD_LIBRARY or ADD_EXECUTABLE.
137 # CUDA_COMPILE_PTX( generated_files file0 file1 ... [OPTIONS ...] )
138 # -- Returns a list of PTX files generated from the input source files.
140 # CUDA_INCLUDE_DIRECTORIES( path0 path1 ... )
141 # -- Sets the directories that should be passed to nvcc
142 # (e.g. nvcc -Ipath0 -Ipath1 ... ). These paths usually contain other .cu
145 # CUDA_WRAP_SRCS ( cuda_target format generated_files file0 file1 ...
146 # [STATIC | SHARED | MODULE] [OPTIONS ...] )
147 # -- This is where all the magic happens. CUDA_ADD_EXECUTABLE,
148 # CUDA_ADD_LIBRARY, CUDA_COMPILE, and CUDA_COMPILE_PTX all call this
149 # function under the hood.
151 # Given the list of files (file0 file1 ... fileN) this macro generates
152 # custom commands that generate either PTX or linkable objects (use "PTX" or
153 # "OBJ" for the format argument to switch). Files that don't end with .cu
154 # or have the HEADER_FILE_ONLY property are ignored.
156 # The arguments passed in after OPTIONS are extra command line options to
157 # give to nvcc. You can also specify per configuration options by
158 # specifying the name of the configuration followed by the options. General
159 # options must preceed configuration specific options. Not all
160 # configurations need to be specified, only the ones provided will be used.
162 # OPTIONS -DFLAG=2 "-DFLAG_OTHER=space in flag"
164 # RELEASE --use_fast_math
165 # RELWITHDEBINFO --use_fast_math;-g
166 # MINSIZEREL --use_fast_math
168 # For certain configurations (namely VS generating object files with
169 # CUDA_ATTACH_VS_BUILD_RULE_TO_CUDA_FILE set to ON), no generated file will
170 # be produced for the given cuda file. This is because when you add the
171 # cuda file to Visual Studio it knows that this file produces an object file
172 # and will link in the resulting object file automatically.
174 # This script will also generate a separate cmake script that is used at
175 # build time to invoke nvcc. This is for several reasons.
177 # 1. nvcc can return negative numbers as return values which confuses
178 # Visual Studio into thinking that the command succeeded. The script now
179 # checks the error codes and produces errors when there was a problem.
181 # 2. nvcc has been known to not delete incomplete results when it
182 # encounters problems. This confuses build systems into thinking the
183 # target was generated when in fact an unusable file exists. The script
184 # now deletes the output files if there was an error.
186 # 3. By putting all the options that affect the build into a file and then
187 # make the build rule dependent on the file, the output files will be
188 # regenerated when the options change.
190 # This script also looks at optional arguments STATIC, SHARED, or MODULE to
191 # determine when to target the object compilation for a shared library.
192 # BUILD_SHARED_LIBS is ignored in CUDA_WRAP_SRCS, but it is respected in
193 # CUDA_ADD_LIBRARY. On some systems special flags are added for building
194 # objects intended for shared libraries. A preprocessor macro,
195 # <target_name>_EXPORTS is defined when a shared library compilation is
198 # Flags passed into add_definitions with -D or /D are passed along to nvcc.
200 # The script defines the following variables:
202 # CUDA_VERSION_MAJOR -- The major version of cuda as reported by nvcc.
203 # CUDA_VERSION_MINOR -- The minor version.
205 # CUDA_VERSION_STRING -- CUDA_VERSION_MAJOR.CUDA_VERSION_MINOR
207 # CUDA_TOOLKIT_ROOT_DIR -- Path to the CUDA Toolkit (defined if not set).
208 # CUDA_SDK_ROOT_DIR -- Path to the CUDA SDK. Use this to find files in the
209 # SDK. This script will not directly support finding
210 # specific libraries or headers, as that isn't
211 # supported by NVIDIA. If you want to change
212 # libraries when the path changes see the
213 # FindCUDA.cmake script for an example of how to clear
214 # these variables. There are also examples of how to
215 # use the CUDA_SDK_ROOT_DIR to locate headers or
216 # libraries, if you so choose (at your own risk).
217 # CUDA_INCLUDE_DIRS -- Include directory for cuda headers. Added automatically
218 # for CUDA_ADD_EXECUTABLE and CUDA_ADD_LIBRARY.
219 # CUDA_LIBRARIES -- Cuda RT library.
220 # CUDA_CUFFT_LIBRARIES -- Device or emulation library for the Cuda FFT
221 # implementation (alternative to:
222 # CUDA_ADD_CUFFT_TO_TARGET macro)
223 # CUDA_CUBLAS_LIBRARIES -- Device or emulation library for the Cuda BLAS
224 # implementation (alterative to:
225 # CUDA_ADD_CUBLAS_TO_TARGET macro).
226 # CUDA_curand_LIBRARY -- CUDA Random Number Generation library.
227 # Only available for CUDA version 3.2+.
228 # CUDA_cusparse_LIBRARY -- CUDA Sparse Matrix library.
229 # Only available for CUDA version 3.2+.
230 # CUDA_npp_LIBRARY -- NVIDIA Performance Primitives library.
231 # Only available for CUDA version 4.0+.
232 # CUDA_nvcuvenc_LIBRARY -- CUDA Video Encoder library.
233 # Only available for CUDA version 3.2+.
235 # CUDA_nvcuvid_LIBRARY -- CUDA Video Decoder library.
236 # Only available for CUDA version 3.2+.
240 # James Bigler, NVIDIA Corp (nvidia.com - jbigler)
241 # Abe Stephens, SCI Institute -- http://www.sci.utah.edu/~abe/FindCuda.html
243 # Copyright (c) 2008 - 2009 NVIDIA Corporation. All rights reserved.
245 # Copyright (c) 2007-2009
246 # Scientific Computing and Imaging Institute, University of Utah
248 # This code is licensed under the MIT License. See the FindCUDA.cmake script
249 # for the text of the license.
253 # License for the specific language governing rights and limitations under
254 # Permission is hereby granted, free of charge, to any person obtaining a
255 # copy of this software and associated documentation files (the "Software"),
256 # to deal in the Software without restriction, including without limitation
257 # the rights to use, copy, modify, merge, publish, distribute, sublicense,
258 # and/or sell copies of the Software, and to permit persons to whom the
259 # Software is furnished to do so, subject to the following conditions:
261 # The above copyright notice and this permission notice shall be included
262 # in all copies or substantial portions of the Software.
264 # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
265 # OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
266 # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
267 # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
268 # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
269 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
270 # DEALINGS IN THE SOFTWARE.
272 ###############################################################################
276 # We need to have at least this version to support the VERSION_LESS argument to 'if' (2.6.2) and unset (2.6.3)
278 cmake_minimum_required(VERSION 2.6.3)
281 # This macro helps us find the location of helper files we will need the full path to
282 macro(CUDA_FIND_HELPER_FILE _name _extension)
283 set(_full_name "${_name}.${_extension}")
284 # CMAKE_CURRENT_LIST_FILE contains the full path to the file currently being
285 # processed. Using this variable, we can pull out the current path, and
286 # provide a way to get access to the other files we need local to here.
287 get_filename_component(CMAKE_CURRENT_LIST_DIR "${CMAKE_CURRENT_LIST_FILE}" PATH)
288 set(CUDA_${_name} "${CMAKE_CURRENT_LIST_DIR}/FindCUDA/${_full_name}")
289 if(NOT EXISTS "${CUDA_${_name}}")
290 set(error_message "${_full_name} not found in ${CMAKE_CURRENT_LIST_DIR}/FindCUDA")
291 if(CUDA_FIND_REQUIRED)
292 message(FATAL_ERROR "${error_message}")
294 if(NOT CUDA_FIND_QUIETLY)
295 message(STATUS "${error_message}")
299 # Set this variable as internal, so the user isn't bugged with it.
300 set(CUDA_${_name} ${CUDA_${_name}} CACHE INTERNAL "Location of ${_full_name}" FORCE)
303 #####################################################################
304 ## CUDA_INCLUDE_NVCC_DEPENDENCIES
307 # So we want to try and include the dependency file if it exists. If
308 # it doesn't exist then we need to create an empty one, so we can
311 # If it does exist, then we need to check to see if all the files it
312 # depends on exist. If they don't then we should clear the dependency
313 # file and regenerate it later. This covers the case where a header
314 # file has disappeared or moved.
316 macro(CUDA_INCLUDE_NVCC_DEPENDENCIES dependency_file)
317 set(CUDA_NVCC_DEPEND)
318 set(CUDA_NVCC_DEPEND_REGENERATE FALSE)
321 # Include the dependency file. Create it first if it doesn't exist . The
322 # INCLUDE puts a dependency that will force CMake to rerun and bring in the
323 # new info when it changes. DO NOT REMOVE THIS (as I did and spent a few
324 # hours figuring out why it didn't work.
325 if(NOT EXISTS ${dependency_file})
326 file(WRITE ${dependency_file} "#FindCUDA.cmake generated file. Do not edit.\n")
328 # Always include this file to force CMake to run again next
329 # invocation and rebuild the dependencies.
330 #message("including dependency_file = ${dependency_file}")
331 include(${dependency_file})
333 # Now we need to verify the existence of all the included files
334 # here. If they aren't there we need to just blank this variable and
335 # make the file regenerate again.
336 # if(DEFINED CUDA_NVCC_DEPEND)
337 # message("CUDA_NVCC_DEPEND set")
339 # message("CUDA_NVCC_DEPEND NOT set")
342 #message("CUDA_NVCC_DEPEND found")
343 foreach(f ${CUDA_NVCC_DEPEND})
344 # message("searching for ${f}")
346 #message("file ${f} not found")
347 set(CUDA_NVCC_DEPEND_REGENERATE TRUE)
351 #message("CUDA_NVCC_DEPEND false")
352 # No dependencies, so regenerate the file.
353 set(CUDA_NVCC_DEPEND_REGENERATE TRUE)
356 #message("CUDA_NVCC_DEPEND_REGENERATE = ${CUDA_NVCC_DEPEND_REGENERATE}")
357 # No incoming dependencies, so we need to generate them. Make the
358 # output depend on the dependency file itself, which should cause the
360 if(CUDA_NVCC_DEPEND_REGENERATE)
361 set(CUDA_NVCC_DEPEND ${dependency_file})
362 #message("Generating an empty dependency_file: ${dependency_file}")
363 file(WRITE ${dependency_file} "#FindCUDA.cmake generated file. Do not edit.\n")
368 ###############################################################################
369 ###############################################################################
370 # Setup variables' defaults
371 ###############################################################################
372 ###############################################################################
374 # Allow the user to specify if the device code is supposed to be 32 or 64 bit.
375 if(CMAKE_SIZEOF_VOID_P EQUAL 8)
376 set(CUDA_64_BIT_DEVICE_CODE_DEFAULT ON)
378 set(CUDA_64_BIT_DEVICE_CODE_DEFAULT OFF)
380 option(CUDA_64_BIT_DEVICE_CODE "Compile device code in 64 bit mode" ${CUDA_64_BIT_DEVICE_CODE_DEFAULT})
382 # Attach the build rule to the source file in VS. This option
383 option(CUDA_ATTACH_VS_BUILD_RULE_TO_CUDA_FILE "Attach the build rule to the CUDA source file. Enable only when the CUDA source file is added to at most one target." ON)
385 # Prints out extra information about the cuda file during compilation
386 option(CUDA_BUILD_CUBIN "Generate and parse .cubin files in Device mode." OFF)
388 # Set whether we are using emulation or device mode.
389 option(CUDA_BUILD_EMULATION "Build in Emulation mode" OFF)
391 # Where to put the generated output.
392 set(CUDA_GENERATED_OUTPUT_DIR "" CACHE PATH "Directory to put all the output files. If blank it will default to the CMAKE_CURRENT_BINARY_DIR")
394 # Parse HOST_COMPILATION mode.
395 option(CUDA_HOST_COMPILATION_CPP "Generated file extension" ON)
397 # Extra user settable flags
398 set(CUDA_NVCC_FLAGS "" CACHE STRING "Semi-colon delimit multiple arguments.")
400 if(CMAKE_GENERATOR MATCHES "Visual Studio")
401 set(CUDA_HOST_COMPILER "$(VCInstallDir)bin" CACHE FILEPATH "Host side compiler used by NVCC")
403 set(CUDA_HOST_COMPILER "${CMAKE_C_COMPILER}" CACHE FILEPATH "Host side compiler used by NVCC")
406 # Propagate the host flags to the host compiler via -Xcompiler
407 option(CUDA_PROPAGATE_HOST_FLAGS "Propage C/CXX_FLAGS and friends to the host compiler via -Xcompile" ON)
409 # Specifies whether the commands used when compiling the .cu file will be printed out.
410 option(CUDA_VERBOSE_BUILD "Print out the commands run while compiling the CUDA source file. With the Makefile generator this defaults to VERBOSE variable specified on the command line, but can be forced on with this option." OFF)
413 CUDA_64_BIT_DEVICE_CODE
414 CUDA_ATTACH_VS_BUILD_RULE_TO_CUDA_FILE
415 CUDA_GENERATED_OUTPUT_DIR
416 CUDA_HOST_COMPILATION_CPP
418 CUDA_PROPAGATE_HOST_FLAGS
421 # Makefile and similar generators don't define CMAKE_CONFIGURATION_TYPES, so we
422 # need to add another entry for the CMAKE_BUILD_TYPE. We also need to add the
423 # standerd set of 4 build types (Debug, MinSizeRel, Release, and RelWithDebInfo)
424 # for completeness. We need run this loop in order to accomodate the addition
425 # of extra configuration types. Duplicate entries will be removed by
427 set(CUDA_configuration_types ${CMAKE_CONFIGURATION_TYPES} ${CMAKE_BUILD_TYPE} Debug MinSizeRel Release RelWithDebInfo)
428 list(REMOVE_DUPLICATES CUDA_configuration_types)
429 foreach(config ${CUDA_configuration_types})
430 string(TOUPPER ${config} config_upper)
431 set(CUDA_NVCC_FLAGS_${config_upper} "" CACHE STRING "Semi-colon delimit multiple arguments.")
432 mark_as_advanced(CUDA_NVCC_FLAGS_${config_upper})
435 ###############################################################################
436 ###############################################################################
437 # Locate CUDA, Set Build Type, etc.
438 ###############################################################################
439 ###############################################################################
441 # Check to see if the CUDA_TOOLKIT_ROOT_DIR and CUDA_SDK_ROOT_DIR have changed,
442 # if they have then clear the cache variables, so that will be detected again.
443 if(NOT "${CUDA_TOOLKIT_ROOT_DIR}" STREQUAL "${CUDA_TOOLKIT_ROOT_DIR_INTERNAL}")
444 unset(CUDA_NVCC_EXECUTABLE CACHE)
445 unset(CUDA_TOOLKIT_INCLUDE CACHE)
446 unset(CUDA_CUDART_LIBRARY CACHE)
447 # Make sure you run this before you unset CUDA_VERSION.
448 if(CUDA_VERSION VERSION_EQUAL "3.0")
449 # This only existed in the 3.0 version of the CUDA toolkit
450 unset(CUDA_CUDARTEMU_LIBRARY CACHE)
452 unset(CUDA_VERSION CACHE)
453 unset(CUDA_CUDA_LIBRARY CACHE)
454 unset(CUDA_cublas_LIBRARY CACHE)
455 unset(CUDA_cublasemu_LIBRARY CACHE)
456 unset(CUDA_cufft_LIBRARY CACHE)
457 unset(CUDA_cufftemu_LIBRARY CACHE)
458 unset(CUDA_curand_LIBRARY CACHE)
459 unset(CUDA_cusparse_LIBRARY CACHE)
460 unset(CUDA_npp_LIBRARY CACHE)
461 unset(CUDA_nvcuvenc_LIBRARY CACHE)
462 unset(CUDA_nvcuvid_LIBRARY CACHE)
465 if(NOT "${CUDA_SDK_ROOT_DIR}" STREQUAL "${CUDA_SDK_ROOT_DIR_INTERNAL}")
466 # No specific variables to catch. Use this kind of code before calling
467 # find_package(CUDA) to clean up any variables that may depend on this path.
469 # unset(MY_SPECIAL_CUDA_SDK_INCLUDE_DIR CACHE)
470 # unset(MY_SPECIAL_CUDA_SDK_LIBRARY CACHE)
473 # Search for the cuda distribution.
474 if(NOT CUDA_TOOLKIT_ROOT_DIR)
476 # Search in the CUDA_BIN_PATH first.
477 find_path(CUDA_TOOLKIT_ROOT_DIR
482 PATH_SUFFIXES bin bin64
483 DOC "Toolkit location."
486 # Now search default paths
487 find_path(CUDA_TOOLKIT_ROOT_DIR
491 DOC "Toolkit location."
494 if (CUDA_TOOLKIT_ROOT_DIR)
495 string(REGEX REPLACE "[/\\\\]?bin[64]*[/\\\\]?$" "" CUDA_TOOLKIT_ROOT_DIR ${CUDA_TOOLKIT_ROOT_DIR})
496 # We need to force this back into the cache.
497 set(CUDA_TOOLKIT_ROOT_DIR ${CUDA_TOOLKIT_ROOT_DIR} CACHE PATH "Toolkit location." FORCE)
499 if (NOT EXISTS ${CUDA_TOOLKIT_ROOT_DIR})
500 if(CUDA_FIND_REQUIRED)
501 message(FATAL_ERROR "Specify CUDA_TOOLKIT_ROOT_DIR")
502 elseif(NOT CUDA_FIND_QUIETLY)
503 message("CUDA_TOOLKIT_ROOT_DIR not found or specified")
508 # CUDA_NVCC_EXECUTABLE
509 find_program(CUDA_NVCC_EXECUTABLE
511 PATHS "${CUDA_TOOLKIT_ROOT_DIR}"
514 PATH_SUFFIXES bin bin64
517 # Search default search paths, after we search our own set of paths.
518 find_program(CUDA_NVCC_EXECUTABLE nvcc)
519 mark_as_advanced(CUDA_NVCC_EXECUTABLE)
521 if(CUDA_NVCC_EXECUTABLE AND NOT CUDA_VERSION)
522 # Compute the version.
523 execute_process (COMMAND ${CUDA_NVCC_EXECUTABLE} "--version" OUTPUT_VARIABLE NVCC_OUT)
524 string(REGEX REPLACE ".*release ([0-9]+)\\.([0-9]+).*" "\\1" CUDA_VERSION_MAJOR ${NVCC_OUT})
525 string(REGEX REPLACE ".*release ([0-9]+)\\.([0-9]+).*" "\\2" CUDA_VERSION_MINOR ${NVCC_OUT})
526 set(CUDA_VERSION "${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR}" CACHE STRING "Version of CUDA as computed from nvcc.")
527 mark_as_advanced(CUDA_VERSION)
529 # Need to set these based off of the cached value
530 string(REGEX REPLACE "([0-9]+)\\.([0-9]+).*" "\\1" CUDA_VERSION_MAJOR "${CUDA_VERSION}")
531 string(REGEX REPLACE "([0-9]+)\\.([0-9]+).*" "\\2" CUDA_VERSION_MINOR "${CUDA_VERSION}")
534 # Always set this convenience variable
535 set(CUDA_VERSION_STRING "${CUDA_VERSION}")
537 # CUDA_TOOLKIT_INCLUDE
538 find_path(CUDA_TOOLKIT_INCLUDE
539 device_functions.h # Header included in toolkit
540 PATHS "${CUDA_TOOLKIT_ROOT_DIR}"
543 PATH_SUFFIXES include
546 # Search default search paths, after we search our own set of paths.
547 find_path(CUDA_TOOLKIT_INCLUDE device_functions.h)
548 mark_as_advanced(CUDA_TOOLKIT_INCLUDE)
550 # Set the user list of include dir to nothing to initialize it.
551 set (CUDA_NVCC_INCLUDE_ARGS_USER "")
552 set (CUDA_INCLUDE_DIRS ${CUDA_TOOLKIT_INCLUDE})
554 macro(FIND_LIBRARY_LOCAL_FIRST _var _names _doc)
555 if(CMAKE_SIZEOF_VOID_P EQUAL 8)
556 # CUDA 3.2+ on Windows moved the library directories, so we need the new
558 set(_cuda_64bit_lib_dir "lib/x64" "lib64" )
560 # CUDA 3.2+ on Windows moved the library directories, so we need to new
561 # (lib/Win32) and the old path (lib).
564 PATHS "${CUDA_TOOLKIT_ROOT_DIR}"
567 PATH_SUFFIXES ${_cuda_64bit_lib_dir} "lib/Win32" "lib"
571 # Search default search paths, after we search our own set of paths.
572 find_library(${_var} NAMES ${_names} DOC ${_doc})
576 find_library_local_first(CUDA_CUDART_LIBRARY cudart "\"cudart\" library")
577 if(CUDA_VERSION VERSION_EQUAL "3.0")
578 # The cudartemu library only existed for the 3.0 version of CUDA.
579 find_library_local_first(CUDA_CUDARTEMU_LIBRARY cudartemu "\"cudartemu\" library")
581 CUDA_CUDARTEMU_LIBRARY
584 # If we are using emulation mode and we found the cudartemu library then use
585 # that one instead of cudart.
586 if(CUDA_BUILD_EMULATION AND CUDA_CUDARTEMU_LIBRARY)
587 set(CUDA_LIBRARIES ${CUDA_CUDARTEMU_LIBRARY})
589 set(CUDA_LIBRARIES ${CUDA_CUDART_LIBRARY})
592 # We need to add the path to cudart to the linker using rpath, since the
593 # library name for the cuda libraries is prepended with @rpath.
594 if(CUDA_BUILD_EMULATION AND CUDA_CUDARTEMU_LIBRARY)
595 get_filename_component(_cuda_path_to_cudart "${CUDA_CUDARTEMU_LIBRARY}" PATH)
597 get_filename_component(_cuda_path_to_cudart "${CUDA_CUDART_LIBRARY}" PATH)
599 if(_cuda_path_to_cudart)
600 list(APPEND CUDA_LIBRARIES -Wl,-rpath "-Wl,${_cuda_path_to_cudart}")
604 # 1.1 toolkit on linux doesn't appear to have a separate library on
606 find_library_local_first(CUDA_CUDA_LIBRARY cuda "\"cuda\" library (older versions only).")
608 # Add cuda library to the link line only if it is found.
609 if (CUDA_CUDA_LIBRARY)
610 set(CUDA_LIBRARIES ${CUDA_LIBRARIES} ${CUDA_CUDA_LIBRARY})
618 #######################
619 # Look for some of the toolkit helper libraries
620 macro(FIND_CUDA_HELPER_LIBS _name)
621 find_library_local_first(CUDA_${_name}_LIBRARY ${_name} "\"${_name}\" library")
622 mark_as_advanced(CUDA_${_name}_LIBRARY)
625 #######################
626 # Disable emulation for v3.1 onward
627 if(CUDA_VERSION VERSION_GREATER "3.0")
628 if(CUDA_BUILD_EMULATION)
629 message(FATAL_ERROR "CUDA_BUILD_EMULATION is not supported in version 3.1 and onwards. You must disable it to proceed. You have version ${CUDA_VERSION}.")
633 # Search for additional CUDA toolkit libraries.
634 if(CUDA_VERSION VERSION_LESS "3.1")
635 # Emulation libraries aren't available in version 3.1 onward.
636 find_cuda_helper_libs(cufftemu)
637 find_cuda_helper_libs(cublasemu)
639 find_cuda_helper_libs(cufft)
640 find_cuda_helper_libs(cublas)
641 if(NOT CUDA_VERSION VERSION_LESS "3.2")
642 # cusparse showed up in version 3.2
643 find_cuda_helper_libs(cusparse)
644 find_cuda_helper_libs(curand)
646 find_cuda_helper_libs(nvcuvenc)
647 find_cuda_helper_libs(nvcuvid)
650 if(NOT CUDA_VERSION VERSION_LESS "4.0")
651 find_cuda_helper_libs(npp)
654 if (CUDA_BUILD_EMULATION)
655 set(CUDA_CUFFT_LIBRARIES ${CUDA_cufftemu_LIBRARY})
656 set(CUDA_CUBLAS_LIBRARIES ${CUDA_cublasemu_LIBRARY})
658 set(CUDA_CUFFT_LIBRARIES ${CUDA_cufft_LIBRARY})
659 set(CUDA_CUBLAS_LIBRARIES ${CUDA_cublas_LIBRARY})
662 ########################
663 # Look for the SDK stuff. As of CUDA 3.0 NVSDKCUDA_ROOT has been replaced with
664 # NVSDKCOMPUTE_ROOT with the old CUDA C contents moved into the C subdirectory
665 find_path(CUDA_SDK_ROOT_DIR common/inc/cutil.h
667 "$ENV{NVSDKCOMPUTE_ROOT}/C"
669 "[HKEY_LOCAL_MACHINE\\SOFTWARE\\NVIDIA Corporation\\Installed Products\\NVIDIA SDK 10\\Compute;InstallDir]"
671 "/Developer/GPU\ Computing/C"
674 # Keep the CUDA_SDK_ROOT_DIR first in order to be able to override the
675 # environment variables.
676 set(CUDA_SDK_SEARCH_PATH
677 "${CUDA_SDK_ROOT_DIR}"
678 "${CUDA_TOOLKIT_ROOT_DIR}/local/NVSDK0.2"
679 "${CUDA_TOOLKIT_ROOT_DIR}/NVSDK0.2"
680 "${CUDA_TOOLKIT_ROOT_DIR}/NV_CUDA_SDK"
681 "$ENV{HOME}/NVIDIA_CUDA_SDK"
682 "$ENV{HOME}/NVIDIA_CUDA_SDK_MACOSX"
686 # Example of how to find an include file from the CUDA_SDK_ROOT_DIR
688 # find_path(CUDA_CUT_INCLUDE_DIR
690 # PATHS ${CUDA_SDK_SEARCH_PATH}
691 # PATH_SUFFIXES "common/inc"
692 # DOC "Location of cutil.h"
695 # # Now search system paths
696 # find_path(CUDA_CUT_INCLUDE_DIR cutil.h DOC "Location of cutil.h")
698 # mark_as_advanced(CUDA_CUT_INCLUDE_DIR)
701 # Example of how to find a library in the CUDA_SDK_ROOT_DIR
703 # # cutil library is called cutil64 for 64 bit builds on windows. We don't want
704 # # to get these confused, so we are setting the name based on the word size of
707 # if(CMAKE_SIZEOF_VOID_P EQUAL 8)
708 # set(cuda_cutil_name cutil64)
710 # set(cuda_cutil_name cutil32)
713 # find_library(CUDA_CUT_LIBRARY
714 # NAMES cutil ${cuda_cutil_name}
715 # PATHS ${CUDA_SDK_SEARCH_PATH}
716 # # The new version of the sdk shows up in common/lib, but the old one is in lib
717 # PATH_SUFFIXES "common/lib" "lib"
718 # DOC "Location of cutil library"
721 # # Now search system paths
722 # find_library(CUDA_CUT_LIBRARY NAMES cutil ${cuda_cutil_name} DOC "Location of cutil library")
723 # mark_as_advanced(CUDA_CUT_LIBRARY)
724 # set(CUDA_CUT_LIBRARIES ${CUDA_CUT_LIBRARY})
728 #############################
729 # Check for required components
732 set(CUDA_TOOLKIT_ROOT_DIR_INTERNAL "${CUDA_TOOLKIT_ROOT_DIR}" CACHE INTERNAL
733 "This is the value of the last time CUDA_TOOLKIT_ROOT_DIR was set successfully." FORCE)
734 set(CUDA_SDK_ROOT_DIR_INTERNAL "${CUDA_SDK_ROOT_DIR}" CACHE INTERNAL
735 "This is the value of the last time CUDA_SDK_ROOT_DIR was set successfully." FORCE)
737 include(${CMAKE_CURRENT_LIST_DIR}/FindPackageHandleStandardArgs.cmake)
738 find_package_handle_standard_args(CUDA
740 CUDA_TOOLKIT_ROOT_DIR
750 ###############################################################################
751 ###############################################################################
753 ###############################################################################
754 ###############################################################################
756 ###############################################################################
757 # Add include directories to pass to the nvcc command.
758 macro(CUDA_INCLUDE_DIRECTORIES)
760 list(APPEND CUDA_NVCC_INCLUDE_ARGS_USER -I${dir})
765 ##############################################################################
766 cuda_find_helper_file(parse_cubin cmake)
767 cuda_find_helper_file(make2cmake cmake)
768 cuda_find_helper_file(run_nvcc cmake)
770 ##############################################################################
771 # Separate the OPTIONS out from the sources
773 macro(CUDA_GET_SOURCES_AND_OPTIONS _sources _cmake_options _options)
775 set( ${_cmake_options} )
777 set( _found_options FALSE )
779 if(arg STREQUAL "OPTIONS")
780 set( _found_options TRUE )
782 arg STREQUAL "WIN32" OR
783 arg STREQUAL "MACOSX_BUNDLE" OR
784 arg STREQUAL "EXCLUDE_FROM_ALL" OR
785 arg STREQUAL "STATIC" OR
786 arg STREQUAL "SHARED" OR
787 arg STREQUAL "MODULE"
789 list(APPEND ${_cmake_options} ${arg})
791 if ( _found_options )
792 list(APPEND ${_options} ${arg})
794 # Assume this is a file
795 list(APPEND ${_sources} ${arg})
801 ##############################################################################
802 # Parse the OPTIONS from ARGN and set the variables prefixed by _option_prefix
804 macro(CUDA_PARSE_NVCC_OPTIONS _option_prefix)
807 # Determine if we are dealing with a perconfiguration flag
808 foreach(config ${CUDA_configuration_types})
809 string(TOUPPER ${config} config_upper)
810 if (arg STREQUAL "${config_upper}")
811 set( _found_config _${arg})
812 # Set arg to nothing to keep it from being processed further
818 list(APPEND ${_option_prefix}${_found_config} "${arg}")
823 ##############################################################################
824 # Helper to add the include directory for CUDA only once
825 function(CUDA_ADD_CUDA_INCLUDE_ONCE)
826 get_directory_property(_include_directories INCLUDE_DIRECTORIES)
828 if(_include_directories)
829 foreach(dir ${_include_directories})
830 if("${dir}" STREQUAL "${CUDA_INCLUDE_DIRS}")
836 include_directories(${CUDA_INCLUDE_DIRS})
840 function(CUDA_BUILD_SHARED_LIBRARY shared_flag)
841 set(cmake_args ${ARGN})
842 # If SHARED, MODULE, or STATIC aren't already in the list of arguments, then
843 # add SHARED or STATIC based on the value of BUILD_SHARED_LIBS.
844 list(FIND cmake_args SHARED _cuda_found_SHARED)
845 list(FIND cmake_args MODULE _cuda_found_MODULE)
846 list(FIND cmake_args STATIC _cuda_found_STATIC)
847 if( _cuda_found_SHARED GREATER -1 OR
848 _cuda_found_MODULE GREATER -1 OR
849 _cuda_found_STATIC GREATER -1)
850 set(_cuda_build_shared_libs)
852 if (BUILD_SHARED_LIBS)
853 set(_cuda_build_shared_libs SHARED)
855 set(_cuda_build_shared_libs STATIC)
858 set(${shared_flag} ${_cuda_build_shared_libs} PARENT_SCOPE)
861 ##############################################################################
862 # Helper to avoid clashes of files with the same basename but different paths.
863 # This doesn't attempt to do exactly what CMake internals do, which is to only
864 # add this path when there is a conflict, since by the time a second collision
865 # in names is detected it's already too late to fix the first one. For
866 # consistency sake the relative path will be added to all files.
867 function(CUDA_COMPUTE_BUILD_PATH path build_path)
868 #message("CUDA_COMPUTE_BUILD_PATH([${path}] ${build_path})")
869 # Only deal with CMake style paths from here on out
870 file(TO_CMAKE_PATH "${path}" bpath)
871 if (IS_ABSOLUTE "${bpath}")
872 # Absolute paths are generally unnessary, especially if something like
873 # file(GLOB_RECURSE) is used to pick up the files.
874 file(RELATIVE_PATH bpath "${CMAKE_CURRENT_SOURCE_DIR}" "${bpath}")
877 # This recipie is from cmLocalGenerator::CreateSafeUniqueObjectFileName in the
881 string(REGEX REPLACE "^[/]+" "" bpath "${bpath}")
882 # Avoid absolute paths by removing ':'
883 string(REPLACE ":" "_" bpath "${bpath}")
884 # Avoid relative paths that go up the tree
885 string(REPLACE "../" "__/" bpath "${bpath}")
887 string(REPLACE " " "_" bpath "${bpath}")
889 # Strip off the filename. I wait until here to do it, since removin the
890 # basename can make a path that looked like path/../basename turn into
891 # path/.. (notice the trailing slash).
892 get_filename_component(bpath "${bpath}" PATH)
894 set(${build_path} "${bpath}" PARENT_SCOPE)
895 #message("${build_path} = ${bpath}")
898 ##############################################################################
899 # This helper macro populates the following variables and setups up custom
900 # commands and targets to invoke the nvcc compiler to generate C or PTX source
901 # dependent upon the format parameter. The compiler is invoked once with -M
902 # to generate a dependency file and a second time with -cuda or -ptx to generate
903 # a .cpp or .ptx file.
905 # cuda_target - Target name
906 # format - PTX or OBJ
907 # FILE1 .. FILEN - The remaining arguments are the sources to be wrapped.
908 # OPTIONS - Extra options to NVCC
910 # generated_files - List of generated files
911 ##############################################################################
912 ##############################################################################
914 macro(CUDA_WRAP_SRCS cuda_target format generated_files)
916 # Set up all the command line flags here, so that they can be overridden on a per target basis.
920 # Emulation if the card isn't present.
921 if (CUDA_BUILD_EMULATION)
923 set(nvcc_flags ${nvcc_flags} --device-emulation -D_DEVICEEMU -g)
925 # Device mode. No flags necessary.
928 if(CUDA_HOST_COMPILATION_CPP)
929 set(CUDA_C_OR_CXX CXX)
931 if(CUDA_VERSION VERSION_LESS "3.0")
932 set(nvcc_flags ${nvcc_flags} --host-compilation C)
934 message(WARNING "--host-compilation flag is deprecated in CUDA version >= 3.0. Removing --host-compilation C flag" )
939 set(generated_extension ${CMAKE_${CUDA_C_OR_CXX}_OUTPUT_EXTENSION})
941 if(CUDA_64_BIT_DEVICE_CODE)
942 set(nvcc_flags ${nvcc_flags} -m64)
944 set(nvcc_flags ${nvcc_flags} -m32)
947 # This needs to be passed in at this stage, because VS needs to fill out the
948 # value of VCInstallDir from within VS. Note that CCBIN is only used if
949 # -ccbin or --compiler-bindir isn't used and CUDA_HOST_COMPILER matches
950 # $(VCInstallDir)/bin.
951 if(CMAKE_GENERATOR MATCHES "Visual Studio")
952 set(ccbin_flags -D "\"CCBIN:PATH=$(VCInstallDir)bin\"" )
955 # Figure out which configure we will use and pass that in as an argument to
956 # the script. We need to defer the decision until compilation time, because
957 # for VS projects we won't know if we are making a debug or release build
959 if(CMAKE_GENERATOR MATCHES "Visual Studio")
960 set( CUDA_build_configuration "$(ConfigurationName)" )
962 set( CUDA_build_configuration "${CMAKE_BUILD_TYPE}")
965 # Initialize our list of includes with the user ones followed by the CUDA system ones.
966 set(CUDA_NVCC_INCLUDE_ARGS ${CUDA_NVCC_INCLUDE_ARGS_USER} "-I${CUDA_INCLUDE_DIRS}")
967 # Get the include directories for this directory and use them for our nvcc command.
968 get_directory_property(CUDA_NVCC_INCLUDE_DIRECTORIES INCLUDE_DIRECTORIES)
969 if(CUDA_NVCC_INCLUDE_DIRECTORIES)
970 foreach(dir ${CUDA_NVCC_INCLUDE_DIRECTORIES})
971 list(APPEND CUDA_NVCC_INCLUDE_ARGS -I${dir})
975 # Reset these variables
976 set(CUDA_WRAP_OPTION_NVCC_FLAGS)
977 foreach(config ${CUDA_configuration_types})
978 string(TOUPPER ${config} config_upper)
979 set(CUDA_WRAP_OPTION_NVCC_FLAGS_${config_upper})
982 CUDA_GET_SOURCES_AND_OPTIONS(_cuda_wrap_sources _cuda_wrap_cmake_options _cuda_wrap_options ${ARGN})
983 CUDA_PARSE_NVCC_OPTIONS(CUDA_WRAP_OPTION_NVCC_FLAGS ${_cuda_wrap_options})
985 # Figure out if we are building a shared library. BUILD_SHARED_LIBS is
986 # respected in CUDA_ADD_LIBRARY.
987 set(_cuda_build_shared_libs FALSE)
989 list(FIND _cuda_wrap_cmake_options SHARED _cuda_found_SHARED)
990 list(FIND _cuda_wrap_cmake_options MODULE _cuda_found_MODULE)
991 if(_cuda_found_SHARED GREATER -1 OR _cuda_found_MODULE GREATER -1)
992 set(_cuda_build_shared_libs TRUE)
995 list(FIND _cuda_wrap_cmake_options STATIC _cuda_found_STATIC)
996 if(_cuda_found_STATIC GREATER -1)
997 set(_cuda_build_shared_libs FALSE)
1001 if(_cuda_build_shared_libs)
1002 # If we are setting up code for a shared library, then we need to add extra flags for
1003 # compiling objects for shared libraries.
1004 set(CUDA_HOST_SHARED_FLAGS ${CMAKE_SHARED_LIBRARY_${CUDA_C_OR_CXX}_FLAGS})
1006 set(CUDA_HOST_SHARED_FLAGS)
1008 # Only add the CMAKE_{C,CXX}_FLAGS if we are propagating host flags. We
1009 # always need to set the SHARED_FLAGS, though.
1010 if(CUDA_PROPAGATE_HOST_FLAGS)
1011 set(_cuda_host_flags "set(CMAKE_HOST_FLAGS ${CMAKE_${CUDA_C_OR_CXX}_FLAGS} ${CUDA_HOST_SHARED_FLAGS})")
1013 set(_cuda_host_flags "set(CMAKE_HOST_FLAGS ${CUDA_HOST_SHARED_FLAGS})")
1016 set(_cuda_nvcc_flags_config "# Build specific configuration flags")
1017 # Loop over all the configuration types to generate appropriate flags for run_nvcc.cmake
1018 foreach(config ${CUDA_configuration_types})
1019 string(TOUPPER ${config} config_upper)
1020 # CMAKE_FLAGS are strings and not lists. By not putting quotes around CMAKE_FLAGS
1021 # we convert the strings to lists (like we want).
1023 if(CUDA_PROPAGATE_HOST_FLAGS)
1024 # nvcc chokes on -g3 in versions previous to 3.0, so replace it with -g
1025 set(_cuda_fix_g3 FALSE)
1027 if(CMAKE_COMPILER_IS_GNUCC)
1028 if (CUDA_VERSION VERSION_LESS "3.0" OR
1029 CUDA_VERSION VERSION_EQUAL "4.1" OR
1030 CUDA_VERSION VERSION_EQUAL "4.2"
1032 set(_cuda_fix_g3 TRUE)
1036 string(REPLACE "-g3" "-g" _cuda_C_FLAGS "${CMAKE_${CUDA_C_OR_CXX}_FLAGS_${config_upper}}")
1038 set(_cuda_C_FLAGS "${CMAKE_${CUDA_C_OR_CXX}_FLAGS_${config_upper}}")
1041 set(_cuda_host_flags "${_cuda_host_flags}\nset(CMAKE_HOST_FLAGS_${config_upper} ${_cuda_C_FLAGS})")
1044 # Note that if we ever want CUDA_NVCC_FLAGS_<CONFIG> to be string (instead of a list
1045 # like it is currently), we can remove the quotes around the
1046 # ${CUDA_NVCC_FLAGS_${config_upper}} variable like the CMAKE_HOST_FLAGS_<CONFIG> variable.
1047 set(_cuda_nvcc_flags_config "${_cuda_nvcc_flags_config}\nset(CUDA_NVCC_FLAGS_${config_upper} ${CUDA_NVCC_FLAGS_${config_upper}} ;; ${CUDA_WRAP_OPTION_NVCC_FLAGS_${config_upper}})")
1050 # Get the list of definitions from the directory property
1051 get_directory_property(CUDA_NVCC_DEFINITIONS COMPILE_DEFINITIONS)
1052 if(CUDA_NVCC_DEFINITIONS)
1053 foreach(_definition ${CUDA_NVCC_DEFINITIONS})
1054 list(APPEND nvcc_flags "-D${_definition}")
1058 if(_cuda_build_shared_libs)
1059 list(APPEND nvcc_flags "-D${cuda_target}_EXPORTS")
1062 # Reset the output variable
1063 set(_cuda_wrap_generated_files "")
1065 # Iterate over the macro arguments and create custom
1066 # commands for all the .cu files.
1067 foreach(file ${ARGN})
1068 # Ignore any file marked as a HEADER_FILE_ONLY
1069 get_source_file_property(_is_header ${file} HEADER_FILE_ONLY)
1070 if(${file} MATCHES ".*\\.cu$" AND NOT _is_header)
1072 # Allow per source file overrides of the format.
1073 get_source_file_property(_cuda_source_format ${file} CUDA_SOURCE_PROPERTY_FORMAT)
1074 if(NOT _cuda_source_format)
1075 set(_cuda_source_format ${format})
1078 if( ${_cuda_source_format} MATCHES "PTX" )
1079 set( compile_to_ptx ON )
1080 elseif( ${_cuda_source_format} MATCHES "OBJ")
1081 set( compile_to_ptx OFF )
1083 message( FATAL_ERROR "Invalid format flag passed to CUDA_WRAP_SRCS for file '${file}': '${_cuda_source_format}'. Use OBJ or PTX.")
1088 # Don't use any of the host compilation flags for PTX targets.
1089 set(CUDA_HOST_FLAGS)
1090 set(CUDA_NVCC_FLAGS_CONFIG)
1092 set(CUDA_HOST_FLAGS ${_cuda_host_flags})
1093 set(CUDA_NVCC_FLAGS_CONFIG ${_cuda_nvcc_flags_config})
1096 # Determine output directory
1097 cuda_compute_build_path("${file}" cuda_build_path)
1098 set(cuda_compile_intermediate_directory "${CMAKE_CURRENT_BINARY_DIR}/CMakeFiles/${cuda_target}.dir/${cuda_build_path}")
1099 if(CUDA_GENERATED_OUTPUT_DIR)
1100 set(cuda_compile_output_dir "${CUDA_GENERATED_OUTPUT_DIR}")
1102 if ( compile_to_ptx )
1103 set(cuda_compile_output_dir "${CMAKE_CURRENT_BINARY_DIR}")
1105 set(cuda_compile_output_dir "${cuda_compile_intermediate_directory}")
1109 # Add a custom target to generate a c or ptx file. ######################
1111 get_filename_component( basename ${file} NAME )
1112 if( compile_to_ptx )
1113 set(generated_file_path "${cuda_compile_output_dir}")
1114 set(generated_file_basename "${cuda_target}_generated_${basename}.ptx")
1115 set(format_flag "-ptx")
1116 file(MAKE_DIRECTORY "${cuda_compile_output_dir}")
1118 set(generated_file_path "${cuda_compile_output_dir}/${CMAKE_CFG_INTDIR}")
1119 set(generated_file_basename "${cuda_target}_generated_${basename}${generated_extension}")
1120 set(format_flag "-c")
1123 # Set all of our file names. Make sure that whatever filenames that have
1124 # generated_file_path in them get passed in through as a command line
1125 # argument, so that the ${CMAKE_CFG_INTDIR} gets expanded at run time
1126 # instead of configure time.
1127 set(generated_file "${generated_file_path}/${generated_file_basename}")
1128 set(cmake_dependency_file "${cuda_compile_intermediate_directory}/${generated_file_basename}.depend")
1129 set(NVCC_generated_dependency_file "${cuda_compile_intermediate_directory}/${generated_file_basename}.NVCC-depend")
1130 set(generated_cubin_file "${generated_file_path}/${generated_file_basename}.cubin.txt")
1131 set(custom_target_script "${cuda_compile_intermediate_directory}/${generated_file_basename}.cmake")
1133 # Setup properties for obj files:
1134 if( NOT compile_to_ptx )
1135 set_source_files_properties("${generated_file}"
1137 EXTERNAL_OBJECT true # This is an object file not to be compiled, but only be linked.
1141 # Don't add CMAKE_CURRENT_SOURCE_DIR if the path is already an absolute path.
1142 get_filename_component(file_path "${file}" PATH)
1143 if(IS_ABSOLUTE "${file_path}")
1144 set(source_file "${file}")
1146 set(source_file "${CMAKE_CURRENT_SOURCE_DIR}/${file}")
1149 # Bring in the dependencies. Creates a variable CUDA_NVCC_DEPEND #######
1150 cuda_include_nvcc_dependencies(${cmake_dependency_file})
1152 # Convience string for output ###########################################
1153 if(CUDA_BUILD_EMULATION)
1154 set(cuda_build_type "Emulation")
1156 set(cuda_build_type "Device")
1159 # Build the NVCC made dependency file ###################################
1160 set(build_cubin OFF)
1161 if ( NOT CUDA_BUILD_EMULATION AND CUDA_BUILD_CUBIN )
1162 if ( NOT compile_to_ptx )
1163 set ( build_cubin ON )
1167 # Configure the build script
1168 configure_file("${CUDA_run_nvcc}" "${custom_target_script}" @ONLY)
1170 # So if a user specifies the same cuda file as input more than once, you
1171 # can have bad things happen with dependencies. Here we check an option
1172 # to see if this is the behavior they want.
1173 if(CUDA_ATTACH_VS_BUILD_RULE_TO_CUDA_FILE)
1174 set(main_dep MAIN_DEPENDENCY ${source_file})
1176 set(main_dep DEPENDS ${source_file})
1179 if(CUDA_VERBOSE_BUILD)
1180 set(verbose_output ON)
1181 elseif(CMAKE_GENERATOR MATCHES "Makefiles")
1182 set(verbose_output "$(VERBOSE)")
1184 set(verbose_output OFF)
1187 # Create up the comment string
1188 file(RELATIVE_PATH generated_file_relative_path "${CMAKE_BINARY_DIR}" "${generated_file}")
1190 set(cuda_build_comment_string "Building NVCC ptx file ${generated_file_relative_path}")
1192 set(cuda_build_comment_string "Building NVCC (${cuda_build_type}) object ${generated_file_relative_path}")
1195 # Build the generated file and dependency file ##########################
1197 OUTPUT ${generated_file}
1198 # These output files depend on the source_file and the contents of cmake_dependency_file
1200 DEPENDS ${CUDA_NVCC_DEPEND}
1201 DEPENDS ${custom_target_script}
1202 # Make sure the output directory exists before trying to write to it.
1203 COMMAND ${CMAKE_COMMAND} -E make_directory "${generated_file_path}"
1204 COMMAND ${CMAKE_COMMAND} ARGS
1205 -D verbose:BOOL=${verbose_output}
1207 -D build_configuration:STRING=${CUDA_build_configuration}
1208 -D "generated_file:STRING=${generated_file}"
1209 -D "generated_cubin_file:STRING=${generated_cubin_file}"
1210 -P "${custom_target_script}"
1211 WORKING_DIRECTORY "${cuda_compile_intermediate_directory}"
1212 COMMENT "${cuda_build_comment_string}"
1215 # Make sure the build system knows the file is generated.
1216 set_source_files_properties(${generated_file} PROPERTIES GENERATED TRUE)
1218 # Don't add the object file to the list of generated files if we are using
1219 # visual studio and we are attaching the build rule to the cuda file. VS
1220 # will add our object file to the linker automatically for us.
1221 set(cuda_add_generated_file TRUE)
1223 if(NOT compile_to_ptx AND CMAKE_GENERATOR MATCHES "Visual Studio" AND CUDA_ATTACH_VS_BUILD_RULE_TO_CUDA_FILE)
1224 # Visual Studio 8 crashes when you close the solution when you don't add the object file.
1225 if(NOT CMAKE_GENERATOR MATCHES "Visual Studio 8")
1226 #message("Not adding ${generated_file}")
1227 set(cuda_add_generated_file FALSE)
1231 if(cuda_add_generated_file)
1232 list(APPEND _cuda_wrap_generated_files ${generated_file})
1235 # Add the other files that we want cmake to clean on a cleanup ##########
1236 list(APPEND CUDA_ADDITIONAL_CLEAN_FILES "${cmake_dependency_file}")
1237 list(REMOVE_DUPLICATES CUDA_ADDITIONAL_CLEAN_FILES)
1238 set(CUDA_ADDITIONAL_CLEAN_FILES ${CUDA_ADDITIONAL_CLEAN_FILES} CACHE INTERNAL "List of intermediate files that are part of the cuda dependency scanning.")
1243 # Set the return parameter
1244 set(${generated_files} ${_cuda_wrap_generated_files})
1248 ###############################################################################
1249 ###############################################################################
1251 ###############################################################################
1252 ###############################################################################
1253 macro(CUDA_ADD_LIBRARY cuda_target)
1255 CUDA_ADD_CUDA_INCLUDE_ONCE()
1257 # Separate the sources from the options
1258 CUDA_GET_SOURCES_AND_OPTIONS(_sources _cmake_options _options ${ARGN})
1259 CUDA_BUILD_SHARED_LIBRARY(_cuda_shared_flag ${ARGN})
1260 # Create custom commands and targets for each file.
1261 CUDA_WRAP_SRCS( ${cuda_target} OBJ _generated_files ${_sources}
1262 ${_cmake_options} ${_cuda_shared_flag}
1263 OPTIONS ${_options} )
1266 add_library(${cuda_target} ${_cmake_options}
1271 target_link_libraries(${cuda_target}
1275 # We need to set the linker language based on what the expected generated file
1276 # would be. CUDA_C_OR_CXX is computed based on CUDA_HOST_COMPILATION_CPP.
1277 set_target_properties(${cuda_target}
1279 LINKER_LANGUAGE ${CUDA_C_OR_CXX}
1285 ###############################################################################
1286 ###############################################################################
1288 ###############################################################################
1289 ###############################################################################
1290 macro(CUDA_ADD_EXECUTABLE cuda_target)
1292 CUDA_ADD_CUDA_INCLUDE_ONCE()
1294 # Separate the sources from the options
1295 CUDA_GET_SOURCES_AND_OPTIONS(_sources _cmake_options _options ${ARGN})
1296 # Create custom commands and targets for each file.
1297 CUDA_WRAP_SRCS( ${cuda_target} OBJ _generated_files ${_sources} OPTIONS ${_options} )
1300 add_executable(${cuda_target} ${_cmake_options}
1305 target_link_libraries(${cuda_target}
1309 # We need to set the linker language based on what the expected generated file
1310 # would be. CUDA_C_OR_CXX is computed based on CUDA_HOST_COMPILATION_CPP.
1311 set_target_properties(${cuda_target}
1313 LINKER_LANGUAGE ${CUDA_C_OR_CXX}
1319 ###############################################################################
1320 ###############################################################################
1322 ###############################################################################
1323 ###############################################################################
1324 macro(CUDA_COMPILE generated_files)
1326 # Separate the sources from the options
1327 CUDA_GET_SOURCES_AND_OPTIONS(_sources _cmake_options _options ${ARGN})
1328 # Create custom commands and targets for each file.
1329 CUDA_WRAP_SRCS( cuda_compile OBJ _generated_files ${_sources} ${_cmake_options}
1330 OPTIONS ${_options} )
1332 set( ${generated_files} ${_generated_files})
1337 ###############################################################################
1338 ###############################################################################
1340 ###############################################################################
1341 ###############################################################################
1342 macro(CUDA_COMPILE_PTX generated_files)
1344 # Separate the sources from the options
1345 CUDA_GET_SOURCES_AND_OPTIONS(_sources _cmake_options _options ${ARGN})
1346 # Create custom commands and targets for each file.
1347 CUDA_WRAP_SRCS( cuda_compile_ptx PTX _generated_files ${_sources} ${_cmake_options}
1348 OPTIONS ${_options} )
1350 set( ${generated_files} ${_generated_files})
1354 ###############################################################################
1355 ###############################################################################
1356 # CUDA ADD CUFFT TO TARGET
1357 ###############################################################################
1358 ###############################################################################
1359 macro(CUDA_ADD_CUFFT_TO_TARGET target)
1360 if (CUDA_BUILD_EMULATION)
1361 target_link_libraries(${target} ${CUDA_cufftemu_LIBRARY})
1363 target_link_libraries(${target} ${CUDA_cufft_LIBRARY})
1367 ###############################################################################
1368 ###############################################################################
1369 # CUDA ADD CUBLAS TO TARGET
1370 ###############################################################################
1371 ###############################################################################
1372 macro(CUDA_ADD_CUBLAS_TO_TARGET target)
1373 if (CUDA_BUILD_EMULATION)
1374 target_link_libraries(${target} ${CUDA_cublasemu_LIBRARY})
1376 target_link_libraries(${target} ${CUDA_cublas_LIBRARY})
1380 ###############################################################################
1381 ###############################################################################
1382 # CUDA BUILD CLEAN TARGET
1383 ###############################################################################
1384 ###############################################################################
1385 macro(CUDA_BUILD_CLEAN_TARGET)
1386 # Call this after you add all your CUDA targets, and you will get a convience
1387 # target. You should also make clean after running this target to get the
1388 # build system to generate all the code again.
1390 set(cuda_clean_target_name clean_cuda_depends)
1391 if (CMAKE_GENERATOR MATCHES "Visual Studio")
1392 string(TOUPPER ${cuda_clean_target_name} cuda_clean_target_name)
1394 add_custom_target(${cuda_clean_target_name}
1395 COMMAND ${CMAKE_COMMAND} -E remove ${CUDA_ADDITIONAL_CLEAN_FILES})
1397 # Clear out the variable, so the next time we configure it will be empty.
1398 # This is useful so that the files won't persist in the list after targets
1399 # have been removed.
1400 set(CUDA_ADDITIONAL_CLEAN_FILES "" CACHE INTERNAL "List of intermediate files that are part of the cuda dependency scanning.")